ETHOnline 2026
Give an AI agent a wallet and it will spend it. Mandate is the part that decides what is worth buying, buys only that, and proves what it paid.
Each route is a whole strategy, priced end to end. A route only qualifies if even its worst case fits the budget, so it can never start something it cannot afford to finish.
A seller saying it was paid is not proof. Mandate checks Hedera's own record of the transaction it signed. Every ID below is real and you can look it up.
The expensive data was never bought for the four quiet pools. That is where the saving comes from.
A third real run. The events seller quotes above the ceiling it published, so Mandate will not pay it. Rather than give up, it re-plans around the refusal and still delivers inside budget.
The same job, smaller budget. Most systems would spend what they had and hand back a half-answer.
A separate harness reads the sellers' own logs and the payment records, then compares them against what Mandate claimed. A test passes only when both agree. It caught a real bug where Mandate reported success while quietly paying twice.
| What is tested | What it proves |
|---|---|
| A reply goes missing after paying | The order is recovered using the original payment. The buyer never pays a second time. |
| The program is restarted mid-job | It picks up what it already bought instead of buying it all again. |
| The budget is too small | The seller is left with no payment at all. Refusing really means refusing. |
git clone https://github.com/mandate-run/mandate
harness/scripts/setup
harness/scripts/demo 1You need two Hedera testnet accounts and a few HBAR. Nothing else: the sellers serve canned pool data, so no API keys are involved. demo 3 shows the refusal, and demo 5 drops a response after payment to show it recover.